The graphics card manufactured by AMD and NVIDIA have certain consumer names like "GeForce GTX 960" or "Radeon R9 380". These follow some loose scheme. The first digit is a series, the remaining digits are what I call a level. There 9 is awesome, 1 is not so good. Taking a trailing zero, one has a level from 10 to 90 or perhaps 95.
Then on the other hand, the chips used in the graphics card are more or less decoupled from this. With AMD, the 380 has the same chip as the 285, it seems to be a mere relabeling. NVIDIA has the 9800 GTX which has the same chip as the GTS 250.
I wondered how the various chips travel through the graphics cards; how the technical names coincide with the marketing names.
First I need a way to normalize the marketing names. With a bunch of regular expressions I could get those down to epoch, series, level and mobile. Since NVIDIA has an "9800 GTX" and a "GTX 980" I use the same "epoch" that are used in Linux distributions when the version numbering scheme has changed. With that I could make a simple normalization:
- "9800 GTX" becomes
(1, 9, 80, False)
- "GTX 980" becomes
(2, 9, 80, False)
- "GTX 1080" becomes
(3, 0, 80, False)
With this, I could add a new number which simply is
epoch * 10 + series. This
would be the name for the card if the manufacturer would have just had one more
This sounds easier than it actually is. I tried to create a flow-chart which allows you to tell which epoch a graphics card belongs to:
Download figure as PDF:
This flowchart has been generated with TikZ,
There are a lot of edge cases which are not included in the diagram. This diagram works with all the cards that I have talked about myself. I for instance have never owned any NVIDIA Quadro or NVS card.
Wikipedia has nice detailed tables which contain the marketing names and the chips that are used in them. Those are HTML and probably really hard to parse. Parsing the Mediawiki syntax might have been a bit easier but it would still be cumbersome. Luckily the Noveau developers have a list which is easier to parse. With a bit of Python I could actually parse most of it.
From there I had a list of chips and which graphics card use it. Using the normalization scheme for the marketing names, I could make a plot like this:
Download figure as PDF:
Each chip has a different color there. It is not really easy to read. Therefore I have also created a grid where each chip is shown. Solid circles are the desktop card, crosses are the mobile versions.
Open the image in a new tab to zoom, its is an SVG image. Or download the PDF:
As one can see, graphics chips are often used in newer cards. No wonder I have a hard time estimating the power a given card will deliver!
Warning: Be careful with those graphics, my regular expressions which parse the marketing names is not perfect. So these graphics might be a bit misleading.
A better ordering of the regular expressions might do the trick. I have included more regular expressions in order to get more data points. At least for some chips (NV34) one can clearly see such a parsing failure.
The code which creates those plots is located here: https://github.com/martin-ueding/graphics-chip-plot